~
Pergamon
Wal. Sci. Tech . Vol. 37. No. II. pp. 73-79 ,1998.
PH: 50273-)223(98)003 )8-7
IAWQ C 1998 Published byElsevier Science LId. Printed in Great Britain. All rights reserved 0273-1223/98 $19'00 + 0'00
TEMPORAL RAINFALL DISAGGREGRATION BASED ON SCALING PROPERTIES J.Olsson* and R. Bemdtsson** * Department o/Civil Engineering, Kyushu University. 6·/0-1
Hakozakl; Hlgashi-ku;
Fukuoka 812, Japan
.*
Department 0/ WaterResources Engineering, Lund University, Box 118, S-221 (){) Lund. Sweden
ABSTRACT The present study concerns disaggregation of daily rainfall time series into higher resolution. For this purpose, the scaling-based cascade model proposed by Olsson (1998) is employed. This model operates by dividing each rainy time period into halves of equal length and distributing the rainfall volume between the halves. For this distribution three possible cases are defined, and the occurrence probability of each case is empiricallyestimated. Olsson (1998) showed that the model was applicable between the time scales I hour and I weekfor rainfall in southern Sweden. Inthe present study, a dailyseasonal (April·June; 3 years)rainfall time series from the same region was disaggregated by the model to 4S-min resolution. The disaggregated data was shown to very well reproduce many fundamental characteristics of the observed 4S-min data, e.g.• the division between rainy and dry periods, the event structure. and the scaling behavior. The results demonstrate the potential of scaling-based approaches in hydrological applications involving rainfall. @ 1998 Published by Elsevier Science Ltd. All rights reserved
KEYWORDS Rainfall; daily timeseries;disaggregation; scaling; cascade. INTRODUCTION The present study focuses on disaggregation of rainfall data into sets of higherresolution. The development of accurate methods for rainfall disaggregation is presently of prime importance in applied hydrology since the solution of hydrological problems often requires rainfall data of a higher resolution than the available measurements. For temporal rainfall, available data is generally of daily resolution whereas hydrological models often require hourly or even finervalues, particularly in urbanapplications. For their calibration, temporal disaggregation models developed to date generally require access to highresolution measurements, i.e., measurements of the same resolution as the model is intended to disaggregate the available data to, from a station nearby the location of interest (e.g., Hershenhom and Woolhiser, 1987), Firstly, such data may not be available. Secondly, due to microclimatological differences it is difficult to judge the representativity of such data for the location of interest (e.g., Econopouly et al., ) 990). To overcome these difficulties, the disaggregation could be basedon scalingproperties of the rainfall data.Then the model could, in principle, be calibrated on the very same data which are to be disaggregated and the uncertainty associated with usingdata fromother locations would thus be eliminated. During the present decade, so-called scaling properties have successively come to be regarded as a fundamental feature of the rainfall process. Scaling in general refers to a statistical symmetry across scales manifesting itself in relationships validover a rangeof scales. Pioneering workin the field was perfonnedby Lovejoy (1982) and later a distinction was made between simple scaling, where the scaling is defined by a 73
74
J. OLSSON and R. BERNDTSSON
single exponent, and multiscaling, where a function is required (e.g., Schertzer and Lovejoy, 1987; Gupta and Waymire, 1990). Scaling and multiscaling properties of rainfall time series have been found in a number of empirical investigations (e.g., Hubert et a/., 1993; Olsson et al., 1993; Olsson, 1995; Burlando and Rosso, 1996; Carsteanu and Foufoula-Georgiou, 1996; Harris et al.; 1996; Svensson et a/., 1996, Tessier et a/., 1996; Menabde et a/., 1997). So-called cascade processes have been proposed as a possible mechanism to account for the scaling properties as well as the hierarchical and clustered structure found in rainfall observations (e.g., Schertzer and Lovejoy, 1987; Gupta and Waymire, 1990, 1993; Tessier et al., 1993; Over and Gupta, 1994). Cascade processes originate from turbulence theory and describe how some quantity is transferred and concentrated from larger to smaller scales in the process (e.g., Yaglom, 1966; Mandelbrot, 1974). The theoretical basis of using cascade processes to model rainfall is yet unclear, but their use is supported mainly by empirical evidence. The relevance of scaling properties for rainfall disaggregation has recently been recognized in some studies. Bo et al. (1994) used the Bartlett-Lewis rectangular pulses to disaggregate daily rainfall into hourly values, and argued that the successful result was due to a scaling (power-law) behavior of the power spectrum. For spatial rainfall, Perica and Foufoula-Georgiou (1996) developed a disaggregation model partly based on scaling of probability distributions of rainfall fluctuations. Olsson (1998) proposed a temporal rainfall disaggregation model based on a conceptually simple cascade scheme. The model showed to be applicable between the approximate time scale limits I hour and 1 week for rainfall in southern Sweden. However, in Olsson (1998) daily values were not disaggregated due to the temporal resolution of the data used. Furthermore, seasonal nonstationarities in the data were not explicitly taken into account, only their magnitude were estimated. Therefore the aim of this paper is to further test the model of Olsson (1998) by performing disaggregation of daily values from a physically homogeneous season.
METHODS
Model The model used is thoroughly described in Olsson (1998), and for detailed information the reader is referred to that paper. Here only a summary of the model is given. Figure 1 illustrates the basic idea of how a conceptually simple cascade scheme was used in Olsson (1998) to represent the temporal structure of rainfall. In the cascade process, a certain time period T associated with a certain rainfall volume V was divided into two halves, T, and Tz, each receiving a part of V, VI = WI'V and Vz=Wz'V, where WI and Wz are multiplicative weights (OsWs I). Three types of division were considered, (1) WI=O and Wz=l, (2) W,=I and Wz=O, and (3) W,=x and Wz=l-x. O
Temporal rainfall disaggregration
7S
Level:
o
100
x/ \x I
45
55
x/V 2
25
0/\1
I
'tv I
20
20
x/ \"
1/\0
35
1/\0
3
4
D . no rain observed
~
•
I)
mm rain observed
Figure I. Principal structure of the cascade scheme used in the present study to represent the temporal rainfall process. with the model formulation (P(x/x)=I). TIme periods T characterized by different positions and volumes were, as assumed, associated with different sets of P(O/l)-, P(1/0)-, and P(x/x)-values .
Rainfall data The data employed originate from a detailed observation program of short-term rainfall properties carried out in the city of Lund, southern Sweden, 1979-1981 (e.g., Niemczynowicz, 1984). The rainfall intensity was measured with a time resolution of I min by small tipping-bucket gauges with an intensity resolution of 0.033 mm/min. The aim of the present study is to disaggregate daily values into successively halved time periods, which means that the resolution of the data must be 24 hours divided by an integer power of 2. Since Olsson (1998) found the present model valid down to at least I hour for similar data, the l-min values are 5 here aggregated into 45-min values (24 hours divided by 2 equals 45 min). In Olsson (1998) it was also revealed that the model parameters, i.e., probability values, showed a slight seasonal nonstationarity. Therefore only spring data (April-June) are used in the present study and the total number of values in the time series is 8192.
Disaggregation As previously mentioned, in the present study the model is used to disaggregate a time series from l-day to 45-min resolution. To achieve this, the probability values P(O/I), P(I/O), and P(x/x) must first be estimated. For this purpose. the weights W were extracted by aggregating the series from 45-min to I-day resolution. In the disaggregation, the mean of the empirical probabilities in the range 45 min to I day was used as probability values. See Olsson (1998) for more detailed information about this parameter estimation
1. OLSSON and R. BERNDTSSON
76
procedure. The model may then be applied for temporal rainfall disaggregation in a straightforward way, as demonstrated by Olsson (1998). Startingfrom the time series at I-day resolution, the position and volumeof each non-zero value is firstly used to determine the set of probability values to be used. Then, for each nonzero value,a randomnumber is drawn to determine the type of division. In case of type 3, a x value is drawn from its corresponding theoretical probability distribution to specify the amount of rainfall in each half. When all non-zero values have been disaggregated accordingly, a series of I2-hour resolution has thus been produced. The procedure was then repeated four times to finally arrive at a series of 45-min resolution. The entire disaggregation from l-day resolution was repeated 10 times, i.e., 10 realizations of the 45-min series was produced. To evaluatethe accuracy of the data generated by the model,these were compared with the observed data in terms of mean and standard deviation of the following five variables: (I) percentage of zero-values, (2) rainfall volume of individual values, (3) rainfall volume of events, (4) duration of events, and (5) length of dry periodsbetween events. At all scales,an event was definedas a sequence of consecutive non-zero values. Furthermore, for all variables but the first quantile-quantile plots were made to evaluatethe agreement of the entire cumulative distribution function. Finallythe ability of the model to reproduce the scaling behaviorof the observed data was studiedby calculating statistical moments of variousordersat different scales. Scaling of moments is manifested in a power law relationship between momentand scale (see, e.g., Svensson et a/., 1996). RESULTS AND DISCUSSION Table 1 shows the overall results from the disaggregation. The model results are obtained as averagesover the 10 realizations produced by the model. From Table I it is evident that the model performs very well in reproducing all the five variablesconsidered. The only pronounced discrepancy is an underestimation of the Table I. Comparison between observeddata (Obs) and data generated by the model(Model), in termsof five variables at all scalesto whichdisaggregation was performed (mean±std).
Zero values Individual volume Eventvolume (mm) (%) (mm)
Eventduration (hrs)
Dryperiod (hrs)
Scale
Data
12 hrs
Obs Model
0.82 0.82
2.9±3.2 2.8±3.5
5.0±6.7 5.2±6.0
21.0±14.2 22.3±14.1
98±94 101±92
6 hrs
Obs Model
0.89 0.88
2.2±2.5 2.1±2.6
3.9±5.2 3.7±4.4
10.4±7.2 10.4±5.9
67±89 65±86
3hrs
Obs Model
0.93 0.93
1.7±2.0 1.7±2.1
3.0±3.6 2.9±3.7
5.2±3.4 5.0±2.9
56±85 52±81
1.5 hrs
Obs Model
0.95 0.95
1.3±1.6 1.311.6
2.3±2.9 2.5±2.9
2.6±1.8 2.7±1.7
52182 55±82
45 min
Obs Model
0.97 0.97
1.0±1.3 1.0±1.2
1.8±2.3 2.0±2.6
1.310.9 1.5±1.0
4l±75 46±77
77
Temporal rainfalldisaggregration
standard deviation of eventvolume andduration at the larger scales. Otherwise the agreement between observed and generated data is nearly total. It is interesting to note that the result does not deteriorate with decreasing scale,as could be suspected beforehand, but also the 45-mindata appearvery well reproduced by the model. Figure2 showstypical examples of quantile-quantile plotscomparing the entire distribution offour variables in the observed and generated 45-min data. Overall the agreement is good, but some differences exist. Individual values(Fig. 2a) largerthan 3-4 mm are generally somewhat overestimated by the model,whereas eventvolumes (Fig.2b) are somewhat underestimated. The lattermaybe due to an excessive tendency of the disaggregation procedure to split up events in two parts at some scale. This also results in a slight underestimation of the event durations (Fig. 2c; note that one point may represent many identical pairs of values). Finally, dry period lengths (Fig. 2d) are well reproduced, which is expected since the main dry
b
a 12
16
----~
r------------,j
14
n
10
E g
8
I .."
C
12 10
'C
J
8 '5 6 E Cii 4 III
6
II
'5
E
4
en
2
n
2
10
468
2
2
12
4
6
Observed (mm)
400
'8
C
7
300
6 't:"
't:"
C
:5.5
J
:5. 'C S 200
4
"
'S
::I
.l;i 3
10 12 14 16
d
C
.!!
8
Observed (mm)
E
C
Cii
tI)
100
2 1 2
3
4
5
Observed (hr)
6
7
8
100
200
300
400
Observed (hr)
Figure2. Quantile-quantile plots comparing the distribution of (a) individual volumes, (b) event volumes, (c) eventdurations, and (d) dry periodlengths in observed and simulated 45-mindata, respectively.
J. OLSSONand R. BERNDTSSON
78
periods in the l-day data are preserved by the model. Figure 3 shows moments of various orders as a function of scale for the observed and generated data. The log-log linear (power-law) curves of the observed data confirm the scaling behavior. The disaggregated data very well match these curves and it is evident that the model preserves the observed scaling behavior. 7.-----------~-~----------------~
8
5
• o
c c
o _ -0'
o ·1
8 days ·2
1 day
--t---- - --t- - -- - - - -- -+-- -- ' Log[scale]
Figure 3. Statistical moments of order h ranging from 0.5 to 4.0 as a function of scale for observed (solid lines) and simulated (point values) 45-min, respectively.
CONCLUSIONS
It is encouraging that the conceptually simple cascade model was able to disaggregate the daily time series into 45-min resolution with many fundamental rainfall characteristics accurately reproduced. This demonstrates the potential of scaling-based approaches in hydrological applications involving rainfall. However, in order for the model to be useful in real-world applications, it should be possible to accurately estimate its parameters using larger-scale data only. This possibility could not be tested in the present study due to the limited amount of data available. The results by Olsson (1998), however, indicated that this can be performed with a high accuracy. In that study, disaggregation was performed using parameter values estimated from time scales larger than approximately I day. The accuracy of the model generated data was similar as compared to using parameter values estimated from the entire scale interval, i.e., down to the resolution to which the disaggregation was performed. Further evaluation of the model using large databases from different geographical regions will be the subject of future research. Generally, more analyses of the scaling properties of rainfall data, not least continuous time series, urgently need to be carried out. In real-world applications of scaling-based hydrological methodologies, such as the present one, an assumption of scaling from the available scale to the desired scale must be made, unless some smaller-scale data are available for the model calibration. Due to the small number of empirical analyses performed to date, such an assumption is at present rather daring, although the few analyses performed nearly unanimously point at the existence of scaling down to small time and space scales. However, more analyses confirming the scaling behavior of rainfall in different geographical regions would substantially increase the confidence in scaling assumptions.
Temporalrainfalldisaggregralion
79
REFERENCES Bo Z., Islam S. and Eltahir EAB. (1994). Aggregation-disaggregation properties of a stochastic rainfall model. Water Resour. Res., 30, 3423-3435. Burlando P. and Rosso R. (1996). Scaling and multiscaling models of depth-duration-frequency curves for storm precipitation. 1. Hydrol., 187,45-64. C&steanu A. and Foufoula-Georgiou E. (1996). Assessing dependence of weights in a multiplicative cascade model of temporalrainfall.1. Geophys. Res., 101, 26363-26370. Econopouly T. W., Davis D. R. and Woolhiser D. A. (1990). Parameter transferability for a daily rainfall disaggregation model. 1. Hydrol., 118, 209·228, 1990. Gupta V. K. and Waymire E. (1990). Multiscaling properties of spatial rainfalland river flow distributions. 1. Geophys. Res., 95, 1999-2009. Gupta V. K. and Waymire E. (1993). A statistical analysisof mesoscalerainfall as a random cascade. 1. Appl. Meteorol., 32, 251-267. Harris D., Menabde M., Seed A. and Austin G. (1996). Multifractal characterization of rain fields with a strong orographic influence. 1. Geophys. Res., 101,26405-26414. Hershenhom J. and Woolhiser D. A. (1987). Disaggregation of daily rainfall. 1. Hydrol., 95, 299-322. Hubert P., Tessier Y., Lovejoy S., Schertzer D., Schmitt F., Ladoy P., Carbonnel J.P., Violette S. and Desurosne I. (1993). Multifractals and extreme rainfallevents. Geophys. Res. Lett., 20, 931-934. LovejoyS. (1982). Area-perimeter relationfor rain and cloud areas. Science, 216, 185-187. Mandelbrot B. B. (1974). Intermittent turbulence in self-similarcascades: divergence of high moments and dimensionof the carrier. 1. Fluid Mech., 62, 331-358. Menabde M., Harris D., Seed A., Austin G. and Stow D. (1997). Multiscaling properties of rainfall and boundedrandomcascades. Water Resour: Res., 33, 2823-2830. Niemczynowicz J. (1984). An investigation ofthe areal and dynamic properties ofrainfall and its influence on runoff generating processes. PhD thesis, Department of Water Resources Engineering, Lund University. Olsson J. (1995). Limits and characteristics of the multifractal behavior of a high-resolution rainfall time series. Nonlinear Proc. Geophys., 2, 23-29. Olsson J. (1998). Evaluation of a scaling cascade model for temporal rainfall disaggregation. Hydrol. Earth System Sci., (in press). Olsson J., Niemczynowicz J. and Bemdtsson R. (1993). Fractal analysis of high-resolution rainfall time series. 1. Geophys. Res., 98, 23265-23274. Over T. M. and Gupta V. K. (1994). Statistical analysis of mesoscale rainfall: dependence of a random cascadegeneratoron large-scale forcing. 1. Appl. Meteorol., 33, 1526-1542. Perica S. and Foufoula-Georgiou E. (1996). Model for multiscale disaggregation of spatial rainfall based on couplingmeteorological and scalingdescriptions. 1. Geophys. Res., 101, 26347-26361. SchertzerD. and Lovejoy S. (1987). Physical modeling and analysisof rain and clouds by anisotropic scaling multiplicative processes. 1. Geophys. Res., 92, 9693-9714. Svensson C., Olsson 1. and Bemdtsson R. (1996). Multifractal properties of daily rainfall in two different climates. Water Resour. Res., 32, 2463-2472. Tessier Y., Lovejoy S. and Schertzer D. (1993). Universal multifractals: theory and observationsfor rain and clouds. 1. Appl. Meteorol., 32, 223-250. TessierY., Lovejoy S., Hubert P., SchertzerD. and Pecknold S. (1996). Multifractal analysisand modelingof rainfall and river flows and scaling,causal transferfunctions. 1. Geophys. Res., 101,26427-26440. Yaglom Y. (1966). The influence of fluctuations in energy dissipation on the shape of turbulence characteristics in the inertial interval. Sov. Phys. Dokl., 11, 26-29.